Testing your assistant as a customer

The best way to know whether your assistant is actually good at selling is to talk to it the way a real customer would — not to read its prompt.

Write the script first, not the questions as you go

Decide the persona and the situation before you open the chat tester:

• Who are they — what kind of business, roughly how big, what channel do their own customers use?
• What are they worried about? A price objection, a bad experience with a chatbot before, a hard “is this even for someone my size”?
• What would make them say yes, and what would make them walk away?

Write down 3–6 things you plan to say, in order, before you type the first one. A script you invent turn by turn tends to drift toward whatever the assistant just said — you end up testing whether it can keep a conversation going, not whether it handles what a real prospect would actually raise.

Run it in the chat tester

Open your dashboard’s chat tester and play the script exactly as written, one message at a time. Do not coach the assistant, do not rephrase a question it misunderstood, and do not skip a beat because the reply already covered it — a real customer would not know to do any of that either.

What to look for

• Did it ask before it assumed? A good assistant confirms what it doesn’t know rather than guessing.
• Did it stay on the numbers you gave it? Watch for a total, a monthly figure or a comparison that doesn’t match what you’d actually be charged — the price should always come from a real quote, never from the assistant doing the maths itself.
• Did it push after you said no, or after you sounded genuinely unsure? It should back off, not lean in.
• Did it name a competitor unfairly, or claim a result it has no way of knowing? Both are against the rules it’s given, and both should never happen.
• Did it ask for your number or email when it didn’t need to — for example when you’d written in from WhatsApp, where it can already reach you?

Write down what you found

For each run, keep the transcript and a short verdict: what went right, what went wrong, and — if something went wrong — what you’d have wanted it to say instead. Two testers on the same script, independently, catch more than one tester running it twice.

What happens to it next

A transcript alone is a story nobody re-checks. What makes a test worth running again is turning what you found into something that can fail loudly the next time:

• A pattern that keeps recurring becomes a support request, so it can be looked at properly rather than described from memory.
• A hardening we make in response to a real test does not go away quietly — it gets a case of its own so the fix stays proven, not just remembered.

This is also how Kooday keeps its own sales assistant honest: this same script lives on as one of the fixed cases it is graded against before any change to its sales prompt ships, so a fix that works today cannot quietly regress later.